diff options
| -rw-r--r-- | usr/src/uts/i86pc/vm/htable.c | 19 |
1 files changed, 17 insertions, 2 deletions
diff --git a/usr/src/uts/i86pc/vm/htable.c b/usr/src/uts/i86pc/vm/htable.c index 1ffb562599..44e66ddfc1 100644 --- a/usr/src/uts/i86pc/vm/htable.c +++ b/usr/src/uts/i86pc/vm/htable.c @@ -571,7 +571,7 @@ htable_steal(uint_t cnt, boolean_t reap) htable_t *list = NULL; htable_t *ht; uint_t stolen = 0; - uint_t pass; + uint_t pass, passes; uint_t threshold; /* @@ -583,11 +583,26 @@ htable_steal(uint_t cnt, boolean_t reap) htable_steal_passes = mmu.ptes_per_table; /* + * If we're stealing merely as part of kmem reaping (versus stealing + * to assure forward progress), we don't want to actually steal any + * active htables. (Stealing active htables merely to give memory + * back to the system can inadvertently kick off an htable crime wave + * as active processes repeatedly steal htables from one another, + * plummeting the system into a kind of HAT lawlessness that can + * become so violent as to impede the one thing that can end it: the + * freeing of memory via ARC reclaim and other means.) So if we're + * reaping, we limit ourselves to the first pass that steals cached + * htables that aren't in use -- which gives memory back, but averts + * the entire breakdown of social order. + */ + passes = reap ? 0 : htable_steal_passes; + + /* * Loop through all user hats. The 1st pass takes cached htables that * aren't in use. The later passes steal by removing mappings, too. */ atomic_inc_32(&htable_dont_cache); - for (pass = 0; pass <= htable_steal_passes && stolen < cnt; ++pass) { + for (pass = 0; pass <= passes && stolen < cnt; ++pass) { threshold = pass * mmu.ptes_per_table / htable_steal_passes; mutex_enter(&hat_list_lock); |